Commodity Feature Extraction In Chinese Online Comments Based On Deep Learning

With the improvement of e-commerce system evaluation system,online shopping comments play a very important role in guiding consumers shop.Online shopping comments objectively reflect consumers' evaluation of shopkeepers' service attitude,after-sales service,product quality,appearance,size and other aspects.Consumers can generally grasp the advantages and disadvantages of commodities and merchants according to the comments,and buy prefer commodities according the comments data.At the same time,merchants can find existing problems in store according to the comments,timely find defects in the commodities,and timely supplement the commodities with good reputations.In this paper,it is necessary to automatically extract the commodity features and corresponding evaluation of commodity features in commodity comments.It can more intuitively reflect the advantages and disadvantages of commodities and user preferences.The extraction of important information in text belongs to the task of information extraction in the field of natural language processing.This paper realizes the extraction of commodity features and their corresponding evaluation in text comments by Deep Learning,so as to avoid manually summarizing complex feature rules.This model is suitable for many kinds of commodity categories.This paper mainly does the following work:(1)Since Self-Attention mechanism can capture the dependence of any absolute position and relative position of text.This paper proposes to add a Self-Attention layer to the Bi-LSTM-CRF information extraction model to compensate the loss of context-dependent information caused by gradient disappearance in the Recurrent Neural Network,and enhance the model's ability to remember contextual information.(2)This paper carry out from the two aspects of character-based char vector and lexical word vector to explore the best way to realize the extraction of commodity features and evaluation information.It can be concluded from the experiment that the character-level and lexical-level models differ little in the recognition of commodity features,but the corresponding evaluation expressions of features are diverse.The performance of character-level model is better than that of lexical-level model,which is mainly due to the variety of feature evaluation methods.The lexical-level model encounters a bottleneck in Chinese word segmentation.(3)The features of commodities are mostly nouns,the corresponding evaluation of features are mostly adjectives.In this context,this paper proposes to convert lexical features into distributed vector representation by binary coding and linear transformation.It plays an important role in improving the recognition rate of nouns and adjectives based on lexical-level model.